DPI-score: A deep learning-based metric for assessing protein-protein interfaces in cryo-EM derived assemblies

Abstract Advances in sample preparation, data acquisition, and data processing have led to a surge of high-resolution cryoEM (cryogenic electron microscopy) derived structures. Although the global resolution continues to improve, local resolution within a map generally varies, and many structures are still resolved at intermediate resolution. Model building and refinement are usually challenging at lower resolutions (>=3 Å), and therefore, atomic model validation is crucial. CryoEM-derived assemblies often contain extensive protein-protein interfaces, yet no established metrics specifically assess the quality of these interfaces. Existing metrics typically focus on fit to density or overall model quality, such as geometry, but do not directly evaluate the protein–protein interfaces in these assemblies. To address this, a Deep learning-based Protein Interface score (DPI-Score) is presented. The proposed approach only requires the raw structural coordinates of the interface atoms without any derived features for training and inference, and achieves 87.53% validation accuracy. The developed interface focused, density independent metric is systematically compared to existing protein interface scoring functions and density dependent model validation scores. It is extensively tested for its performance to distinguish ‘target-like’ interfaces from a pool of predicted models (cryoEM oligomeric targets from CASP15 and CASP16) and score fitted structures generated using rigid body fitting. We further applied it to assess 29,120 interfaces from the 6,011 fitted entries (worse than 3 Å) associated with depositions in the Electron Microscopy Data Bank. DPI-score provides complementary information to other scores and is able to detect errors at the interfaces in modelled assemblies that were not identified using density-based scores alone. The source code for model validation is freely available for academic use on Git-Lab at https://gitlab.com/ccpem/dpi or via the CCP-EM software suite, Doppio, at https://www.ccpem.ac.uk/software .